• classification
• convolutional encoders
• edge computing
• EfficientNet
• feature extraction
• lightweight models
• MobileNetV2
• real-time inference

Point-of-Care Ultrasound (POCUS) is increasingly adopted in emergency and bedside settings, yet its reliability remains constrained by the strong operator dependency of image acquisition. The absence of standardized probe positioning leads to heterogeneous image quality, limiting both human interpretation and the performance of AI-based diagnostic tools.
This thesis investigates the feasibility of real-time view verification directly on constrained mobile devices, using retrospective B-mode ultrasound data as the basis for model design and evaluation. A lightweight pipeline was developed, combining preprocessing, feature extraction through convolutional encoders, dimensionality reduction, and classification of frame acceptability.
Experiments focused on suprapatellar knee ultrasound and compared the proposed models with state-of-the-art baselines. The encoder classifier achieved 94.75% accuracy on the test set (precision 95.94%, recall 94.50%, ROC–AUC ≈0.99), while EfficientNet and MobileNetV2 obtained 99.25% and 98.00% accuracy, respectively.
However, baseline models proved roughly 30 times slower even on GPU, making them unsuitable for real-time use. In contrast, the lightweight design sustained over 5FPS on constrained devices, with CPU latencies below 200ms and sub-millisecond inference on GPU, thereby meeting the requirements for clinical usability.
Overall, the thesis demonstrates that mobile, on-device view verification is both technically feasible and clinically meaningful, providing a concrete step toward reducing operator dependence and enabling standardized POCUS acquisition in real-world practice.